I was working in a busy GP clinic when I finished seeing my last patient for the day, and I felt a sense of relief that the day was over; however, that relief was quickly overshadowed by the daunting mountain of paperwork I needed to complete before I could go home. In general practice, this work pattern is becoming expected practice throughout the profession. GPs routinely spend 1 to 3 hours outside work daily, catching up on patient notes, referrals, insurance claims, and lab results. This pattern is becoming more untenable for many, contributing to the increasing rates of burnout, mental health issues, and intention to quit being higher than they have ever been in modern medicine. Although other aspects of medicine have been progressing, there has been little headway made with regard to clinician efficiency and satisfaction in the workplace. Until now…

Over the last six months, I have used a “generative-conversational artificial intelligence” (AI scribe for short) to help me with my job. The AI scribe writes all of my clinic notes just by listening to my consultation. It seems like magic compared to other digital tools we use in medicine, and I feel like I can do the job I was trained to do rather than an expensive record keeper. It takes the self-imposed and managerial time pressure away, significantly improving rapport with my patients. It feels like you are participating in a practical/OSCE exam, where you are doing your best because everything you say is transcribed and summarized. What patients find impressive are the improved engagement with clinicians and the ability to leave the consultation with a well-worded and written treatment plan. In some literature, they have been shown to foster improved engagement with patients by improving communication and understanding and reducing clinician multitasking, which is known to increase the rates of misdiagnoses and errors.

So how does it work? AI scribes are mostly browser-based and use the same processes: listening to the conversation, performing a voice-to-text translation, generating a transcript, and then sending the transcript for summarizing. AI scribes like Nabla copilot summarize the transcript using ChatGPT 3’s underlying parent AI, GPT3. Following a predetermined template, this summary creates a clinical note for the patient’s health record. Additional notes and dictations can be added as needed, and the note (output) can be copied and pasted into the patient record or even printed to give to the patient. Some can rewrite the note as a referral to a particular specialty and include ICD-10 coding.

Obtaining patient consent is of the utmost importance when utilizing the AI scribe. This ensures that patients are fully aware of how the transcription and summary process is conducted and how their data is used. In my experience, the reception to this has been overwhelmingly positive over the last six months, with only one patient declining the use of the AI scribe. Most, if not all, appreciated having my undivided attention during the consultation, which further underscores the importance of patient consent in this process.

However, they are far from perfect. There are concerns about these AI scribes regarding information privacy, data standards, and safety. Although companies that own these AIs outline their safety regulatory practices (GDPR and HIPAA compliant), clinicians must be wary of any company that submits patient data to the underlying AI model for reinforcement/teaching the AI. AI needs billions of pieces of text data to learn. Suppose a patient is identified within the transcript. In that case, the potential for a breach of privacy is higher as someone (not in the clinical context) could retrieve information on that consultation by asking for the AI prompts, thereby breaching patient privacy.

There are also issues regarding the workflow that I have found to be archetypal to all AI scribes. If the voice transcription is having difficulty, the note will be sparse and lack meaningful salient points. Moreover, if there are connection or server issues, entire parts of the consultation will be missing.

If transcription misinterpretations occur, these misinterpretations are perpetuated throughout the document, which can be a pain to edit and can impact the diagnostic coding.

I also found inconsistencies between the AI scribes, with the most serious related to patients with multiple problems (especially if they were mental health-related). I found that the detail required for documenting mental health assessments was less than optimal. In a small number of instances, the scribe didn’t generate any of the mental health information or action plan but retained only the physical issues.

AI scribes are browser-based, so they store previous consultation transcripts and summaries locally and on the cloud. Therefore, we must be aware of patient privacy concerns and ensure that you delete all previously generated notes as they are saved on your account to reduce privacy breaches.

Despite the concerns and challenges, it’s important to recognize the potential of AI scribes. They are tools that can free up cognitive space and reduce chronic multitasking, which is known to improve decision-making. This potential for improved efficiency and decision-making is a significant step forward in the evolution of health care practices, and I hope it only continues to get better. I suspect clinicians are already considering how they perform their roles and deal with the mounting responsibilities and pressures. This is the beginning of an equalizer between meeting the clinician’s needs and improving rapport and engagement within busy health services, which need a touch of humanity in today’s chaotic environment.

Jamie Ioane is a general practitioner in New Zealand.